253 	s32 workspace1[8], workspace2[8];  in fwht()  local
317 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
318 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
319 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
320 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
322 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
323 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
324 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
325 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
328 		out[0] = workspace2[0] + workspace2[4];  in fwht()
329 		out[1] = workspace2[0] - workspace2[4];  in fwht()
330 		out[2] = workspace2[1] - workspace2[5];  in fwht()
331 		out[3] = workspace2[1] + workspace2[5];  in fwht()
332 		out[4] = workspace2[2] + workspace2[6];  in fwht()
333 		out[5] = workspace2[2] - workspace2[6];  in fwht()
334 		out[6] = workspace2[3] - workspace2[7];  in fwht()
335 		out[7] = workspace2[3] + workspace2[7];  in fwht()
355 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
356 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
357 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
358 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
360 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
361 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
362 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
363 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
365 		out[0 * 8] = workspace2[0] + workspace2[4];  in fwht()
366 		out[1 * 8] = workspace2[0] - workspace2[4];  in fwht()
367 		out[2 * 8] = workspace2[1] - workspace2[5];  in fwht()
368 		out[3 * 8] = workspace2[1] + workspace2[5];  in fwht()
369 		out[4 * 8] = workspace2[2] + workspace2[6];  in fwht()
370 		out[5 * 8] = workspace2[2] - workspace2[6];  in fwht()
371 		out[6 * 8] = workspace2[3] - workspace2[7];  in fwht()
372 		out[7 * 8] = workspace2[3] + workspace2[7];  in fwht()
386 	s32 workspace1[8], workspace2[8];  in fwht16()  local
406 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
407 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
408 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
409 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
411 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
412 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
413 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
414 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
417 		out[0] = workspace2[0] + workspace2[4];  in fwht16()
418 		out[1] = workspace2[0] - workspace2[4];  in fwht16()
419 		out[2] = workspace2[1] - workspace2[5];  in fwht16()
420 		out[3] = workspace2[1] + workspace2[5];  in fwht16()
421 		out[4] = workspace2[2] + workspace2[6];  in fwht16()
422 		out[5] = workspace2[2] - workspace2[6];  in fwht16()
423 		out[6] = workspace2[3] - workspace2[7];  in fwht16()
424 		out[7] = workspace2[3] + workspace2[7];  in fwht16()
444 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
445 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
446 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
447 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
449 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
450 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
451 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
452 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
455 		out[0*8] = workspace2[0] + workspace2[4];  in fwht16()
456 		out[1*8] = workspace2[0] - workspace2[4];  in fwht16()
457 		out[2*8] = workspace2[1] - workspace2[5];  in fwht16()
458 		out[3*8] = workspace2[1] + workspace2[5];  in fwht16()
459 		out[4*8] = workspace2[2] + workspace2[6];  in fwht16()
460 		out[5*8] = workspace2[2] - workspace2[6];  in fwht16()
461 		out[6*8] = workspace2[3] - workspace2[7];  in fwht16()
462 		out[7*8] = workspace2[3] + workspace2[7];  in fwht16()
473 	int workspace1[8], workspace2[8];  in ifwht()  local
494 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
495 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
496 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
497 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
499 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
500 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
501 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
502 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()
505 		out[0] = workspace2[0] + workspace2[4];  in ifwht()
506 		out[1] = workspace2[0] - workspace2[4];  in ifwht()
507 		out[2] = workspace2[1] - workspace2[5];  in ifwht()
508 		out[3] = workspace2[1] + workspace2[5];  in ifwht()
509 		out[4] = workspace2[2] + workspace2[6];  in ifwht()
510 		out[5] = workspace2[2] - workspace2[6];  in ifwht()
511 		out[6] = workspace2[3] - workspace2[7];  in ifwht()
512 		out[7] = workspace2[3] + workspace2[7];  in ifwht()
532 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
533 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
534 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
535 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
537 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
538 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
539 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
540 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()
546 			out[0 * 8] = workspace2[0] + workspace2[4];  in ifwht()
547 			out[1 * 8] = workspace2[0] - workspace2[4];  in ifwht()
548 			out[2 * 8] = workspace2[1] - workspace2[5];  in ifwht()
549 			out[3 * 8] = workspace2[1] + workspace2[5];  in ifwht()
550 			out[4 * 8] = workspace2[2] + workspace2[6];  in ifwht()
551 			out[5 * 8] = workspace2[2] - workspace2[6];  in ifwht()
552 			out[6 * 8] = workspace2[3] - workspace2[7];  in ifwht()
553 			out[7 * 8] = workspace2[3] + workspace2[7];  in ifwht()
560 			out[0 * 8] = workspace2[0] + workspace2[4];  in ifwht()
561 			out[1 * 8] = workspace2[0] - workspace2[4];  in ifwht()
562 			out[2 * 8] = workspace2[1] - workspace2[5];  in ifwht()
563 			out[3 * 8] = workspace2[1] + workspace2[5];  in ifwht()
564 			out[4 * 8] = workspace2[2] + workspace2[6];  in ifwht()
565 			out[5 * 8] = workspace2[2] - workspace2[6];  in ifwht()
566 			out[6 * 8] = workspace2[3] - workspace2[7];  in ifwht()
567 			out[7 * 8] = workspace2[3] + workspace2[7];  in ifwht()